SoftCache: Adaptive Step Size Caching for Diffusion Transformers
Diffusion models are slow because each step requires a full network pass. SoftCache improves this by using 'headroom'—the distance from a staleness threshold—to dynamically adjust the step size instead of just deciding to refresh or reuse computation. This adaptive jumping allows the sampler to skip scheduled nodes, leading to faster generation and superior fidelity across various models.